original version presented in S. Riniker, G. Landrum, J. Cheminf., 5, 26 (2013), DOI: 10.1186/1758-2946-5-26, URL: http://www.jcheminf.com/content/5/1/26
extended version presented in S. Riniker, N. Fechner, G. Landrum, J. Chem. Inf. Model., 53, 2829, (2013), DOI: 10.1021/ci400466r, URL: http://pubs.acs.org/doi/abs/10.1021/ci400466r
The virtual-screening process implemented by the benchmarking platform is divided into three steps:
-
Scoring
-
Validation
-
Analysis
The three steps are run separately and read in the output of the previous step. In the scoring step, the data from the directories compounds and query_lists is read in.
The directory compounds contains lists of compounds for 118 targets from three public data sources: MUV, DUD and ChEMBL. The compound lists contain the external ID, the internal ID and the SMILES of each compound.
There are three subsets of targets available:
subset I: 88 targets from MUV, DUD & ChEMBL described in J. Cheminf., 5, 26 (2013)
subset I filtered: 69 targets from MUV, DUD & ChEMBL filtered for difficulty described in JCIM (2013), online
subset II: 37 targets from ChEMBL designed for a second VS use case described in JCIM (2013), online
The directory query_lists contains training lists for each target with the indices of randomly selected active and inactive molecules. Training lists with 5, 10 or 20 active molecules are available. The number of training decoys is 20 % of the decoys for subsets I and 10 % for subset II.
The scripts are written in Python and use the open-source cheminformatics library RDKit (www.rdkit.org) and machine-learning library scikit-learn (www.scikit-learn.org).
Running a script with the option [--help] gives a description of the required and optional input parameters of the script.